Best probe direction is not the best steering direction, Fraunhofer study finds across models

FraunhoferIAIS · hf · 2026-11-20

Fraunhofer IAIS researchers systematically compare probing and steering directions in language models' hidden states. Constructing a one-parameter family of directions from a covariance-adjusted mean difference, they evaluate probing and additive steering at a fixed side effect. For truthfulness and sentiment, in every tested instruction-tuned model, the best probe is not the best steering direction — steering consistently favors directions closer to the mean difference, and the separation holds across 59 of 60 cases, with implications for transfer and abstention.

Original post →

More from Research

Research channel →